Journal of Computing and Information Technology - CIT 13, 2005, 3, 211–223
211
Arabic Font Recognition using Decision
Trees Built from Common Words
Ibrahim S. I. Abuhaiba
Department of Electrical and Computer Engineering, Islamic University of Gaza, Gaza, Palestine
We present an algorithm for a priori Arabic optical Font
Recognition (AFR). The basic idea is to recognize fonts
of some common Arabic words. Once these fonts are
known, they can be generalized to lines, paragraphs, or
neighbor non-common words since these components of
a textual material almost have the same font. A decision
tree is our approach to recognize Arabic fonts. A set
of 48 features is used to learn the tree. These features
include horizontal projections, Walsh coefficients, invariant moments, and geometrical attributes. A set of 36
fonts is investigated. The overall success rate is 90.8%.
Some fonts show 100% success rate. The average time
required to recognize the word font is approximately
0.30 seconds.
Keywords: Arabic fonts, optical font recognition,vertical
normalization, projections, Walsh coefficients, invariant
moments, geometrical features, decision tree learning.
1. 1. Introduction
Optical font recognition can be addressed
through two complementary approaches: the a
priori approach, in which characters of the analyzed text are not yet known, and the a posteriori
approach, where the content of the given text is
used to recognize the font. To our knowledge,
there has been nearly no studies of the Arabic
Font Recognition (AFR) problem. Available
studies deal with Latin fonts, which have different characteristics than Arabic fonts. Therefore,
in this paper, we present a novel solution to the
a priori AFR problem.
Often, the font style is not the same for a whole
document; it is a word feature, rather than a
document feature, and its detection can be used
to discriminate between different regions of the
document, such as title, figure caption, or normal text. The detection of the font style of
a word can also be used to improve character recognition: we know that Mono-font OCR
systems achieve better results than Multi-font
ones, so the recognition of a document can be
done using first an OFR, and then a Mono-font
OCR.
To our knowledge, the only work that addresses
the AFR problem is that of reference 1 ], where
an algorithm for AFR is presented. First, words
in the training set of documents for each font
are segmented into symbols that are rescaled.
Next, templates are constructed, where every
new training symbol that is not similar to existing templates is considered as a new template.
Templates are sharable between fonts. To classify the font of a word, its symbols are matched
to the templates and the fonts of the best matching templates are retained. The most frequent
font is the word font. The overall error, rejection
and success rates were 15.0%, 7.6%, and 77.4%,
respectively. The high error rate is mainly due
to errors in size recognition.
In 2 ], font recognition is developed to enhance
the recognition accuracy of a text recognition
system. Font information is extracted from two
sources: one is the global page properties and
the other is the graph matching result of recognized short words such as a, it, and of.
In 3 ], a multi-font OCR system to be used for
document processing is presented. The system performs, at the same time, both character
recognition and font-style detection of the digits belonging to a subset of the existing fonts.
The detection of the font-style of the document
words can guide a rough automatic classification of documents, and can also be used to improve character recognition. The system uses
212
Arabic Font Recognition using Decision Trees Built from Common Words
the tangent distance as a classification function
in a nearest neighbour approach. The nearest
neighbour approach is always able to recognize
the digit, but the performance in font detection
is not optimal. To improve the performance,
they used a discriminant model, the TD-Neuron,
which is used to discriminate between two similar classes.
In 6 ], a study of image degradations effects on
the performance of a font recognition system is
presented. The evaluation that has been carried
out shows that the system is robust against natural degradations such as those introduced by
scanning and photocopying, but its performance
decreases with very degraded document images.
In order to avoid this weakness, a degradation
modeling strategy has been adapted, allowing
an automatic adaptation of the system to these
degradations. The adaptation is derived from
statistical analysis of features behavior against
degradations and is performed by specific transformations applied to the system knowledge
base.
In 4 ], a texture analysis-based approach is used
for font recognition. Existing methods are typically based on local features that often require
connected components analysis. In this work,
the document is taken as an image containing
some special textures, and font recognition as
texture identification. The method is content
independent and involves no local feature analysis. The well-established 2-D Gabor filtering technique is applied to extract such features
and a weighted Euclidean distance classifier is
used in the recognition task. The reported average recognition accuracy of 24 fonts over 6,000
samples is 98.6%.
All previous OFR studies, except the first, deal
with Latin fonts and there have been no similar
studies on Arabic fonts, which have different
characteristics than Latin fonts. The most impeding characteristic of Arabic OCR systems is
the cursive nature of Arabic script, which makes
characters not ready for direct OCR or OFR. In
this paper, we present a novel contribution to
the a priori AFR problem.
In 5 ], a statistical approach based on global typographical features is proposed for font recognition. It aims at the identification of the typeface, weight, slope, and size of the text from an
image block without any knowledge of the content of that text. The recognition is based on a
multivariate Bayesian classifier and operates on
a given set of known fonts. The effectiveness of
the adopted approach has been experimented on
a set of 280 fonts. Font recognition accuracies
of about 97% are reached on high-quality images. Rates higher than 99.9% were obtained for
weight and slope detection. Experiments have
also shown the system robustness to document
language and text content and its sensitivity to
text length.
Due to difficulties in segmenting cursive Arabic
text, the basic idea in our method is to recognize
fonts of some common Arabic words. Once
these fonts are known, they can be generalized
to lines, paragraphs, or neighbor non-common
words since these components of a textual material almost have a single font. Figure 1 shows
one line of Arabic text written in three different
fonts, from top to bottom: Simplified Arabic,
Traditional Arabic, and Tahoma. This line contains the words “
”, “
”, and “ ”, which
are all common Arabic words used in all contexts irrespective of the subject of material presented. If we could recognize the font of these
a)
b)
c)
Fig. 1. One line of Arabic text written in three different fonts, all Roman and Regular, from top to bottom: (a)
Simplified Arabic, (b) Traditional Arabic, and (c) Tahoma.
Arabic Font Recognition using Decision Trees Built from Common Words
words only, it can be assigned to the rest of
words in the line.
Decision tree learning is one of the most widely
used and practical methods for inductive inference. It is a method for approximating functions that is robust to noisy data. So, a decision tree is our approach to recognize Arabic
fonts. In our AFR system, a font is identified
by four attributes: typeface (Simplified Arabic, Traditional Arabic, etc.), size expressed in
typographic points, slant (Roman, Italic), and
weight (Regular, Bold).
The rest of the paper is organized as follows.
Common Arabic words are selected in Section
2. The feature set to be used in the font classification decision tree is presented in Section
3. The learning algorithm of the font decision
tree is described in Section 4. The recognition
process is explained in Section 5. In Section 6,
experimental results are reported. Finally, the
paper is concluded in Section 7.
2. Finding Common Arabic Words
Font recognition of every character is a very difficult and time-consuming problem, especially
in cursive scripts such as Arabic. The reason of
this difficulty is that words have to be segmented
first into characters, which is also another difficult problem and no efficient solution to the
213
segmentation problem is available. Actually,
the details of character segmentation depend on
the word font which, in current OCR systems,
remains unknown until the characters are recognized. Therefore, we chose not to detect the
font per character.
In document analysis systems, the stage of layout analysis results in segmenting the page into
logical components. Examples of such components are non-textual components such as
graphs, photos, line arts, etc., and textual components such as titles, section headings, and
paragraphs. When we talk about font recognition, we are interested in textual components.
Assuming that a textual component is written in
almost one font, it is not necessary to detect the
font of every word in a textual component. It is
sufficient to detect the font of some representative words and generalize that font to the whole
textual component, which can be faster.
The set of representative words are the most
common words used in Arabic texts irrespective of the subject. For easier segmentation, we
selected words that do not contain a separation
character in the middle, see Figure 2 for a list of
these characters. These are characters which,
Fig. 2. Arabic separation characters.
Fig. 3. The 100 most common Arabic words not having a separation character in the middle.
214
Arabic Font Recognition using Decision Trees Built from Common Words
when exist in the middle of the word, separate
it into more than one non-connected sub-words.
Such discontinuity causes difficulties when segmenting lines into complete words. A word
without separation characters is connected and
it is almost very easy to segment it successfully.
So, we compiled very long files of Arabic texts
from many disciplines such as Islamic publications, arts, engineering, science, journalism,
etc. The frequency of every such word is computed. The top 100 words having the highest
frequencies were selected, see Figure 3. These
words were printed in different fonts to constitute the learning dataset and one testing dataset.
Details of the learning and testing datasets can
be found Section 6.
3. Feature Set
1
0
for foreground pixels
for background pixels
(1)
In the following, there is a complete description
of our set of features.
After vertically normalizing a word Image,
I (x y ), its horizontal projection, h(y ), is calculated using the equation:
X
x
Since Walsh transform proved to be powerful
in shape discrimination, we used the 1-D Walsh
discrete transform of the horizontal projections,
h(y ), described above, to find N Walsh coefficients. This provides N other features, fi ,
i = N N + 1 : : : 2N ; 1. Walsh coefficients
are found using the equation:
I (x y )
1
N
;
X
N 1
y =0
;
Y
n 1
h(y )
i=0
;1)b (y)b ; ; (u)
i
(
n 1 i
(3)
where N = 2n , and bk (z ) is the kth bit in the
binary representation of z. We don’t use Walsh
coefficients for the vertical projections since
we perform vertical normalization resulting in a
varying number of columns from word to word.
In practice, Equation (3) is not used directly.
Instead, the fast Walsh transform is found using
the successive doubling method in a way similar
to the FFT 7 ]. Thus, Equation (3) becomes:
W (u) =
1
Weven (u) + Wodd (u)]
2
(4)
and
W (u + M ) =
where M
=
1
Weven (u) ; Wodd (u)]
2
N=2, u = 0 1
:::
(5)
M ; 1.
3.3. Invariant Moments
For a 2-D discrete word image, I (x y ), the moment of order ( p + q) is defined as
3.1. Projections
h(y ) =
3.2. Walsh Coefficients
W (u) =
Some features to be described in this section
need words to be normalized first in order to
get a fixed number of features, which facilitates comparison processes. In the context of
document understanding, the normalization operation almost means to normalize in two directions: x and y. Actually, this can be problematic
since information is lost due to this kind of twodimensional normalization. In our system, we
perform a one-dimensional normalization, in
the vertical direction, such that the height/width
ratio is preserved, which results in normalized
words that retain the relative geometrical attributes of the original un-normalized words.
Assume that the image of a word is enclosed by
a minimum bounding rectangle and is defined
by the function:
I (x y ) =
where N is the word height after normalization.
This yields N features, fi , i = 0 1 : : : N ; 1.
We don’t use the vertical projection since the
number of columns varies from word to word,
which produces different number of features for
different words.
y
=
0 1
:::
N ; 1 (2)
m pq =
XX
x
x py q I (x y )
(6)
y
The central moments can be expressed as
µ pq =
XX
x
y
(x
; x)p(y ; y)qI (x
y)
(7)
215
Arabic Font Recognition using Decision Trees Built from Common Words
where
x=
m10
m00
(8)
and
m01
m00
The central moments of order up to 3 are:
y=
(9)
µ00 = m00
µ10 = 0
µ01 = 0
µ20 = m20 ; x m10
µ02 = m02 ; y m01
µ11 = m11 ; y m10
µ30 = m30 ; 3x m20 + 2x2 m10
µ12 = m12 ; 2y m11 ; x m02 + 2y2 m10
µ21 = m21 ; 2x m11 ; y m20 + 2x2 m01
µ03 = m03 ; 3y m02 + 2y2 m01
The normalized central moments, denoted, η pq,
are defined as
µ pq
η pq = γ
(10)
µ00
where
γ =
p+q
+ 1 for
2
p+q = 2 3
:::
φ2 = (η20 ;
η02 )2 +
(12)
2
4η11
φ3 = (η30 ; η12 )2 + (3η21 ; η03 )2
φ4 = (η30 + η12 )2 + (η21 + η03 )2
(13)
(14)
(15)
φ5 = (η30 ; 3η12 )(η30 + η12 )
2
(η 30 + η 12 )
; 3(η21 + η03)2 ]
; η03)(η21 + η03)
2
2
3(η 30 + η 12 ) ; (η 21 + η 03 ) ]
+ (3η 21
(16)
2
(η 21 +η 03 ) ]
φ6 = (η20 ;η02 )(η30 +η12 ) ;
+ 4η 11 (η 30 + η 12 )(η 21 + η 03 )
2
φ7 = (3η21 ; η30 )(η30 + η12 )
2
(η 30 + η 12 )
(17)
; 3(η21 + η03)2 ]
; η30)(η21 + η03)
2
2
3(η 30 + η 12 ) ; (η 21 + η 03 ) ]
+ (3η 12
3.4. Geometrical Features
These include: width, height, aspect ratio, area,
x-coordinate of center of area, y-coordinate of
center of area, perimeter, thinness ratio, and
direction of axis of least second moment. After word segmentation, the limits, x min , x max ,
y min , and y max , of the bounding rectangle become known. The width, D, and height, H, of a
word are defined as:
D = x max ; x min + 1
(19)
; ymin + 1
(20)
and
H
= y max
Width and height features provide information
about the size of the bounding rectangle of the
word. The aspect ratio, α , is defined by
(11)
In our font recognition system, we used the following seven invariant moments which are derived from the second and third moments:
φ1 = η20 + η02
This set of moments, which will be denoted by
fi , i = 2N 2N + 1 : : : 2N + 6, is invariant to
translation, rotation, and scale change 7 ].
(18)
D
H
α=
(21)
The area, A, which equals the number of 1 pixels in the image, indicates the relative size of
the word, and is defined by:
A=
XX
x
I (x y )
(22)
y
Note that the area equals m00 defined earlier.
The x and y coordinates of the center of area
are the parameters x and y defined previously.
These two features locate where in the bounding rectangle the center of area lies. To find
the perimeter, P, for every 1 pixel we find the
number of 0 pixels in its 4-neighbors. The sum
of these numbers for all 1 pixels in the word
yields the perimeter, P. The thinness ratio, T,
is calculated by
T
=
4π
A
P2
(23)
As the perimeter becomes larger relative to the
area, this ratio decreases and the word gets
thinner. This can be useful in discriminating
between Regular and Bold fonts and between
thin and thick fonts. The axis of least second
216
Arabic Font Recognition using Decision Trees Built from Common Words
moment feature provides information about the
word’s orientation, which can be helpful in discriminating between Roman and Italic fonts.
This axis corresponds to the line about which
it takes the least amount of energy to spin the
word or the axis of least inertia. If we move
our origin to the center of area (x y), the angle,
θ , of axis of least second moment is defined as
follows:
P P xyI (x y )
1
x y
P P x 2I (x
θ = tan ;1 P P 2
2
y I (x y ) ;
2
x
y
x
y
y)
(24)
where θ is measured in counterclockwise direction from the vertical axis.
Geometrical features provide 9 more features
and will be denoted by fi , i = 2N + 7 2N +
8 : : : 2N + 15. Therefore, the total number of
features becomes N horizontal projections + N
Walsh coefficients + 7 invariant moments + 9
geometrical features = 2N + 16 features. In our
system, N was set to 16 yielding 48 total feature
fi = 0 1 : : : 47.
4. Decision Tree Learning
Decision tree learning is one of the most widely
used and practical methods for inductive inference. It is a method for approximating functions
that is robust to noisy data and capable of learning disjunctive expressions 8 ].
Decision trees classify instances by sorting them
down the tree from the root to some leaf node,
which provides the classification of the instance.
Each node in the tree specifies a test of some feature of the instance, and each branch descending
from that node corresponds to one of the possible values for this feature.
Most algorithms that have been developed for
learning decision trees are variations on a core
algorithm that employs a top-down, greedy
search through the space of possible decision
trees 8-10 ]. We will use an algorithm that learns
the decision tree by constructing it top-down beginning with the question “which feature should
be tested at the root of the tree?” To answer this
question, each instance feature is evaluated using a statistical test to determine how well it
alone classifies the training examples. The best
feature is selected and used as the test at the root
node of the tree. A descendent of the root node
is then created for each possible suitable range
of this feature, and the training examples are
sorted to the appropriate descendent node (i.e.,
down the branch corresponding to the example’s range for this feature). The entire process
is then repeated using the training examples associated with each descendent node to select the
best feature to test at that point in the tree. This
forms a greedy search for an acceptable decision
tree, in which the algorithm never backtracks to
reconsider earlier choices.
The central choice in the algorithm is selecting
which feature to test at each node in the tree. We
would like to select the feature that is most useful for classifying examples. There are many
several quantitative measures of the worth of a
feature 8 ]. We will use a statistical property,
called information gain that measures how well
a given feature separates the learning examples
according to their target font classification. The
algorithm uses this measure to select among the
candidate features at each step while growing
the tree.
In order to define information gain precisely, we
begin by defining a measure commonly used in
information theory, called entropy, which characterizes the impurity of an arbitrary collection
of examples. Given a collection S of font examples where the target font classification can take
on c different values, the entropy of S relative
to this c-wise classification is defined as:
Entropy (S) =
c
X
i=1
; pi log2 pi
(25)
where pi is the proportion of S belonging to font
class i.
Given the entropy as a measure of the impurity
in a collection of training examples, we can now
define a measure of the effectiveness of a feature in classifying the training examples. The
measure we will use, called information gain, is
simply the expected reduction in entropy caused
by partitioning the examples according to this
feature. More precisely, the information gain,
Gain(S f ) of a feature f relative to a collection
Arabic Font Recognition using Decision Trees Built from Common Words
of examples S, is defined as
Gain(S f ) Entropy (S)
;
P
jSvj Entropy (S )
v
v2Values( f ) jSj
(26)
where values( f ) is the set of all possible values for feature f , and Sv is the subset of S
for which feature f has value v (i.e., Sv =
fs 2 S j f (s) = vg ). Note that the first term
in Equation (26) is just the entropy of the original collection, S, and the second term is the
expected value of the entropy after S is partitioned using feature f . The expected entropy
described by this second term is simply the sum
of the entropies of each subset, Sv , weighted by
jSvj that belong to S .
the fraction of examples
v
jSj
Gain(S f ) is therefore the expected reduction in
entropy caused by knowing the value of feature
f.
To give a formal description of our font decision tree learning algorithm, let S = fsi i =
0 1 : : : n ; 1g be a set of training examples,
i.e., instance word images, where n is the total
number of examples and each instance, si , is a
2-tuple (V τ ), V = (v0 v1 : : : v47 ), where vj is
the value of feature fj , j = 0 1 : : : 47, and τ is
the value of an additional Font feature, i.e., the
font classification of this example. Notice that
the Font feature is the target function value to be
learned and then predicted by the tree. Initially,
the set S contains the whole training examples.
A node in the tree to be created contains the
following fields:
1. feature: which is used by interior nodes to
store the number of the feature to be tested
at that node,
2. threshold: which is used by interior nodes
to decide which branch to follow. If the
value of the feature to be tested at that node
is less than threshold, then the left branch
is followed, otherwise, the right branch is
followed.
3. font: which is used by leaf nodes to store the
target font classification.
4. Examples: which is used by a leaf node to
store the examples sorted to that node. Here,
we store only the V vector, i.e., the 48 feature
vector values are stored. The font classification per example is not stored since it is
217
common to all examples sorted to this node
and it is retained in the font field.
5. left and right: which are pointers used by an
interior node to point to left and right child
nodes of that node, respectively.
A formal description of our font decision tree
learning algorithm follows.
Algorithm I. Font Decision Tree Learning
(FDTL)
DecisionTreeNode FDTL(S)
Step 1. New Node Creation
Create a Root node for the tree.
Step 2. Leaf Node Check
If all Examples have the same font classification, τ , then
f
g
Mark Root as a leaf node
Let Root ! f ont = τ
Store the V vector, i.e., the 48 feature values
of S in Root ! S
Return Root
Step 3. Best Classifying Feature Selection
Find the feature, fbest , of the features f0 to f47
which best classifies the set S and the corresponding dividing threshold, t best .
Step 4. Interior Node and Descendent Nodes
Creation
Mark Root as an interior node
Set Root ! feature = fbest and
Root ! threshold = t best
Let S<tbest S be the set of examples for which
v fbest < t best
Let Root ! left = FDTL(S<tbest )
Let Stbest S be the set of examples for which
v fbest t best
Let Root ! right = FDTL(Stbest )
Return Root
End of Algorithm I.
Steps 3 of Algorithm I needs explanation. Notice that the feature set, f0 to f47 , in our problem,
218
Arabic Font Recognition using Decision Trees Built from Common Words
is continuous-valued. The decision tree creation
can be made easier if these features are converted into discrete-valued features. Actually,
this is accomplished by dynamically defining
new discrete-valued features that partition the
continuous-valued features into a discrete set of
intervals. In particular, for a feature fj that is
continuous-valued, the algorithm creates a new
Boolean feature b that is true if the value of fj
is less than a certain threshold t and false otherwise. We pick a threshold, t, which produces the
greatest information gain of feature b. This can
be achieved by sorting the examples according
to the continuous feature, fj , then identifying
adjacent examples that differ in their target font
classification, we can generate a set of candidate
thresholds midway between the corresponding
values of fj . The value of t that maximizes
information gain of feature b must lie at such
boundary 11 ]. These candidate thresholds are
evaluated by computing the information gain
associated with each and the best feature can
be selected. This is repeated for every feature,
fj , and the feature which yields the maximum
information gain is selected.
Therefore, in Step 3 of Algorithm I, the best
feature, fbest , which best classifies the set S and
the corresponding dividing threshold, tbest , are
found as follows. For every feature, fj , in the
48 features:
1. Let Y = fy i i = 0 1 : : : n ; 1g, where
y i = (vi j τi ) and vi j is the jth feature value
in si .
2. Sort Y in ascending order according to feature fj
3. Let gainMax
=
0
4. In the following For loop, instead of comparing the font classifications of adjacent examples, we compare font classifications of
y i and y i+δ to speed up the learning phase,
since otherwise a great number of candidate
thresholds arises, which highly increases the
learning time. This is achieved by defining
δ = blog2 nc instead of δ = 1, where n is
the number of examples in S:
For (i = 0; i < n ; δ ; i = i + δ )
f
Let τ i and τ i+δ be the font classes of yi
and y i+δ , respectively.
If τi 6= τi+δ then
f
g
g
t = (vi j + vi+δ j )=2
Define the set of examples Z = fzi =
(bi τ i )g, where τi is the font class in
y i = (vi j τi ) and bi equals 1 if vi j < t
and 0 otherwise.
Let g be equal to the information gain
of feature b relative to the set Z
If g > gainMax then set gainMax =
g and tbest = t
The feature, fbest , which achieves the maximum
information gain and the corresponding threshold, t best are retained.
Pages of common Arabic words, defined in Section 2, are printed and scanned to constitute the
learning dataset. Some image preprocessing is
required in both the learning and testing stages.
First, an image is skew-corrected using the algorithm of 12 ]. Next, horizontal and vertical
solid lines are removed. Third, pepper noise is
removed. The image is segmented into lines using horizontal white cuts, where each line consists of a sequence of words. Then, every line
is segmented into words using vertical white
cuts. Words which don’t satisfy certain size
constraints are filtered out. For example, if the
height or width of a word is less than some specified thresholds, or is greater than some other
thresholds, then the word is discarded, see Section 6 for values of these thresholds. This process is repeated for every image of the training
dataset. Features of extracted words are calculated to obtain the complete set of training
examples. The range of each feature, to be used
in the recognition stage, is found. A formal
description of the complete learning algorithm
follows.
Algorithm II.
Arabic Font Learning (AFL)
DecisionTreeNode AFL()
Step 1. Preprocessing
a. Correct the skew of the input image
b. Remove horizontal and vertical solid lines
c. Remove pepper noise
d. Segment the image into lines using horizontal white cuts
219
Arabic Font Recognition using Decision Trees Built from Common Words
e. Every line is segmented into words using
vertical white cuts
f. Repeat steps (a-e) for all images of the training dataset
Step 2. Feature Calculation
a. Find the 48 feature values of every extracted
word passing some tests to be mentioned in
Section 6. At the end of this step, a set
of training example words, S = fsi i =
0 1 : : : n ; 1g is obtained, where n is the
total number of examples and each instance,
si , is a 2-tuple (V τ ), V = (v0 v2 : : : v47),
where vj is the value of feature fj , j =
0 1 : : : 47, and τ is the value of the additional Font feature or the font classification
of this example.
b. Find the full range, rj , of every feature, fj ,
j
=
0 1
:::
47, where rj
;
i=0
n 1
=
;
i=0
n 1
max(vi j)
;
min(vi j), where vi j is the value of feature fj
in training example si .
Step 3. Decision Tree Learning
Apply the decision tree learning algorithm, Algorithm I, to the set of examples S to obtain the
target font recognition decision tree. Let Root
point to the root of this tree.
Return Root
End of Algorithm II.
5. Recognition
The same preprocessing operations used in the
learning stage are also used in the recognition
stage. However, it is not necessary that all examples extracted in the recognition stage belong
to the set of common Arabic words defined in
Section 2. To recognize the font of a word, its
features are calculated. Feature testing starts at
the root of the tree. Let the feature which must
be tested at the root be f eaturej and the corresponding threshold be thresholdj . The value vj
of feature fj of the word is compared against
thresholdj . If vj < thresholdj then the left
branch is followed, otherwise, the right branch
is followed and the testing process is repeated
at new interior nodes we arrive at. If a leaf node
is reached, a check is made to make sure that
the word being tested belongs to the set of common Arabic words defined in Section 2. This
is achieved by finding whether the distance between the word being tested and any training
example, in the set of training examples found
at that leaf node, is not greater than a previously defined distance threshold. If so, the font
saved at the leaf node is returned as the recognized font. Otherwise, the word is rejected. We
define our distance threshold as:
DistThreshold
=
α
Nf
(27)
where α is some constant that is empirically
determined and N f is the number of features,
which is 48 in our case. The distance measure we incorporated in our test is the rangenormalized city block distance metric defined
as
47 X
vj ; uj (28)
D(V U ) rj j=0
where vj is the value of feature fj in a training
example found in the leaf node and uj is the
corresponding feature value of the word being
tested. This metric is computationally faster
than the Euclidean distance, but gives similar
results. The following is a formal description
of the recognition algorithm per word. Before
we call the algorithm, the set of feature values,
U = (u0 u1 : : : u47 ), of the word to be tested
is calculated.
Algorithm III. Arabic Font Recognition
(AFR)
font AFR(DecisionTreeNode U )
f
If DecisionTreeNode is an interior node,
then
f
g
Let f = DecisionTreeNode ! f eature
If (u f < DecisionTreeNode ! threshold),
then
Return AFR(DecisionTreeNode !
lef t )
Else
Return AFR(DecisionTreeNode !
right )
Else
220
f
g
g
Arabic Font Recognition using Decision Trees Built from Common Words
Learning Dataset
If the range-normalized city block distance between U and any of the examples in the set DecisionTreeNode !
Examples is not greater than DistThreshold, then
Return DecisionTreeNode ! f ont
Else
Return rejected
Here rejected is a dedicated font value
to indicate an unknown font value.
End of Algorithm III.
Words whose fonts are recognized using the
above algorithm can be used to predict the fonts
of constituent headlines, text lines, and paragraphs of a page in many ways. For example, following a maximum voting strategy, a
line/paragraph can be assigned the most frequent font in the set of recognized words in that
line/paragraph. Or, a rejected word can be assigned the font of the nearest recognized word.
6. Results
In our system, a font, τ , is characterized by
the 4-tuple ( p z s w ), where p, z, s, w are the
typeface, size, slant, and weight, respectively.
In the fonts used in experimentation, we have
three typefaces: Simplified Arabic, Traditional
Arabic, and Tahoma. The slant is either Roman
or Italic. The weight is either Regular or Bold.
The size is 12, 13, or 14 points. Thus, a total of
36 fonts were investigated. Table I summarizes
the fonts used in our AFR system.
Typeface
Size
The most frequent 100 Arabic words found in
Section 2, see Figure 3, were printed once for
every font in Table I, i.e., they were printed 3
typefaces 3 point sizes 2 slants 2 weights
= 36 times. In every font, 20 samples were
printed for each word. Thus, the total number
of learning examples is 36 fonts 20 examples per word 100 words per font = 72,000
examples.
Testing Datasets
A dataset, dataset no. 1, similar to that of the
learning dataset was created, but this time 10
samples per word were created. This provided
36,000 samples for testing purposes. Another
dataset, dataset no. 2, was created by compiling a file of normal Arabic text, i.e., it is not
constrained to only common Arabic words of
Figure 3. This file contained 15 paragraphs
and, when printed in the fonts of Table I, yields
from two to three A4 pages per font.
The learning and testing datasets were printed
using a laser jet printer. Then, they were scanned
as binary images with resolution set to 300
dpi in both directions, using an HP ScanJet
3400C scanner. Figure 4 shows one sample
of a scanned paragraph from testing dataset no.
2.
In learning and testing stages, a word that does
not satisfy any of the following constraints is filtered out: the word width and height are at least
12 pixels each, the maximum width and maximum height are 600 and 200 pixels, respectively. These values were empirically determined and proved adequate to eliminate flecks
and some non-textual content.
It is worth mentioning that our font recognition
decision tree is not designed to recognize every Arabic word. It only recognizes instances
of words which belong to the set of common
Arabic words defined in Section 2. If a word is
Slant
Weight
Simplified Arabic
12, 13, 14 Roman, Italic Regular, Bold
Traditional Arabic
12, 13, 14 Roman, Italic Regular, Bold
Tahoma
12, 13, 14 Roman, Italic Regular, Bold
Table I. Arabic fonts used in our AFR system.
Arabic Font Recognition using Decision Trees Built from Common Words
221
Fig. 4. Sample scanned paragraph of the testing dataset no. 2, printed in (Simplified Arabic, Roman, Regular).
outside that set, then it is rejected by the tree.
Therefore, in defining the recognition rate, R,
we don’t depend on the usual definition of this
rate which is:
R=
No: of words correctly recognized
Total no: of words
100
(29)
This is due to the fact that in usual Arabic text
most words don’t belong to our set of common
words, hence, they are rejected by the tree which
gives the incorrect illusion that our algorithm is
not practical. Instead, we use the following definition:
R0 =
No: of words correctly recognized
Total no: of words;No: of rejected words
100
(30)
The recognition algorithm was run with the
threshold DistThreshold set to 0:01 N f =
0:01 48 = 0:48, where N f is the number of
features. For this threshold value, 46.6% of the
total words in dataset no.1 were rejected by the
tree although the training and testing datasets
look similar. However, in font recognition, very
small degradations in intensity values of image
pixels produce new variants of word images that
no longer pass the threshold test of the decision tree. The rejection rate can be decreased
by increasing DistThreshold. However, this increases, as empirically verified, the error rate.
Actually, there is very little harm of the rejection rate being high as long there is an adequate
number of common words successfully recognized. Also, it is not necessary to recognize
every common Arabic word in a paragraph. For
the set of words accepted by the decision tree,
the overall success rate of testing dataset no. 1
is 93.9%. Error is mainly due to errors in size
and slant recognition. Errors in typeface are almost absent, while very few errors in weight are
noticed. Simplified Arabic and Traditional Arabic fonts show high size, slant, and overall error
rates, high rejection rate, and low recognition
success rate compared to Tahoma fonts.
For testing dataset no. 2, 89.5% of input words
were rejected by the decision tree, which is
higher than the rejection rate of dataset no. 1.
The reason is that dataset no. 2 contains too
many other classes of words other than the common words on which the decision tree is trained.
Of accepted words, the overall success rate of
testing dataset no. 2 is 90.8%. Some fonts show
100% success rate. Error is mainly due to errors
in size in contrast to results of dataset no. 1, in
which error was mainly due to errors in size and
slant recognition. Also, errors in typeface and
weight are almost absent. Some Traditional
Arabic fonts show high size and overall error
rates and low recognition success rate compared
to Simplified Arabic and Tahoma fonts. The error rate can be decreased by increasing the size
of the training dataset.
The algorithm was implemented and run on a
Pentium III 866MHz PC with 128 MB RAM.
The average time required to recognize the word
font is approximately 0.31 and 0.29 seconds for
testing datasets nos. 1 and 2, respectively. This
time doesn’t include disk overhead and preprocessing times as these times are counted towards
the overall document understanding software
and not towards every constituent task in that
software. The recognition time can be reduced
by using some programming optimization techniques and more powerful computers.
The size of the learned decision tree is 15633
nodes, 7816 of which are interior nodes and
222
Arabic Font Recognition using Decision Trees Built from Common Words
the rest are leaves. In each interior node, one
feature is tested and the same feature may be
used more than once in a path from the root to a
leaf node. Here, we have the following important observations about the usage of features in
interior nodes of the decision tree:
1. All defined features were used in the tree.
The mostly used feature is the horizontal
projection of row no. 0, h(0), of the vertically-normalized word image, which was used
in 657 nodes out of 7816 interior nodes. The
least used feature is Walsh coefficient no. 9,
W (9), which was used in 30 interior nodes.
2. Most of the heavily used features are very
simple features which are not computationally expensive such as horizontal projections, width, area, and height.
3. Also, most of the heavily used features are
the invariant moments. However, these are
computationally expensive.
Table II summarizes feature usage which shows
that horizontal projections occupy 47.1% of the
interior nodes. This is a useful result since it
indicates that we can depend on simple features
to classify fonts. Next come invariant moments
Feature Name
and Walsh coefficients which constitute 19.3%
and 12.7%, respectively, of the features tested
at interior nodes.
7. Conclusion
In this paper, we presented a novel contribution
to the a priori AFR problem. A decision tree
was our approach to recognize Arabic fonts. A
set of 48 features was used to learn the tree.
These features include horizontal projections,
Walsh coefficients, invariant moments, and geometrical attributes.
A set of 36 fonts was investigated. The total number of learning examples was 72,000.
Two datasets were used to test the method. The
overall success rate is 90.8%. Some fonts show
100% success rate. The error rate can be decreased by increasing the size of the training
dataset. Recognized fonts of common words
can be generalized to their containing paragraphs, assuming that each paragraph contains
only a single font, which can be practical in
printed materials such as newspapers and old
books that were not previously stored in electronic format.
Symbol
Nodes
Percent
h(0) to h(15)
3685
47.1
Invariant moments
φ1 to φ7
1509
19.3
Walsh coefficients
W (0) to W (15)
996
12.7
Width
D
354
4.5
Area
A
265
3.4
Height
H
233
3.0
Center of area x coordinate
x
220
2.8
Center of area y coordinate
y
129
1.7
Perimeter
P
117
1.5
Angle of axis of least second moment
θ
109
1.4
Aspect ratio
α
103
1.3
Thinness ratio
T
96
1.2
7816
100
Horizontal projections
All features
Table II. Summary of feature usage in interior nodes of learned font decision tree. Nodes: no. of interior nodes using
the corresponding feature, Percent: percentage of interior nodes using the feature.
Arabic Font Recognition using Decision Trees Built from Common Words
The algorithm was implemented and run on a
Pentium III 866MHz PC with 128 MB RAM.
The average time required to recognize the word
font is approximately 0.30 seconds. The time
can be reduced by using some programming optimization techniques and more powerful computers.
Regarding feature usage in the decision tree,
horizontal projections occupy 47.1% of the interior nodes. This is a useful result since it
indicates that we can depend on simple features
to classify fonts. Next come invariant moments
and Walsh coefficients which constitute 19.3%
and 12.7%, respectively, of the features tested
at interior nodes.
Acknowledgement
The author thanks the referees for their valuable
comments which helped to improve this paper.
223
8]
T.M. MITCHELL, Machine learning, The McGrawHill Companies, Inc., Singapore, 1997.
9]
J.R. QUINLAN, Induction of Decision Trees, Machine Learning, 1 (1986), pp. 81–106.
10 ]
J.R. QUINLAN,C4.5: Programs for machine learning, San Mateo, CA: Morgan Kaufmann, 1993.
11 ]
U.M. FAYYAD AND K.B. IRANI, On the handling of
continuous-valued attributes in decision tree generation, Machine Learning, 8 (1992), pp. 87–102.
12 ]
I.S.I. ABUHAIBA, Skew Correction of Textual Documents, King Saud University Journal: Computer
and Information Sciences Division, 15 (2003), pp.
67–86.
Received: September, 2004
Revised: December, 2004
Accepted: February, 2005
Contact address:
Ibrahim S. I. Abuhaiba
P. O. Box 1276
Department of Electrical and Computer Engineering
Islamic University of Gaza
Gaza
Palestine
[email protected]
References
1]
I.S.I. ABUHAIBA, Arabic Font Recognition Based
on Templates, The International Arab Journal of
Information Technology, 1 (2003), pp. 33–39.
2]
H. SHI AND T. PAVLIDIS, Font Recognition and Contextual Processing for More Accurate Text Recognition, Forth International Conference on Document
Analysis and Recognition (ICDAR’97), (1997) Ulm,
Germany.
3]
S.L. MANNA, A.M. COLLA, AND A. SPERDUTI, Optical Font Recognition for Multi-Font OCR and
Document Processing, 10th International Workshop
on Database & Expert Systems Applications, (1999)
Florence, Italy.
4]
Y. ZHU, T. TAN, AND Y. WANG, Font Recognition
Based on Global Texture Analysis, Fifth International Conference on Document Analysis and
Recognition (ICDAR’99), (1999) Bangalore, India.
5]
A. ZRAMDINI AND R. INGOLD, Optical Font Recognition Using Typographical Features, IEEE Transactions on Pattern Analysis and Machine Intelligence, 20 (1998), pp. 877–882.
6]
A. ZRAMDINI AND R. INGOLD, A Study of Document
Image Degradation Effects on Font Recognition,
Third International Conference on Document Analysis and Recognition (ICDAR’95), (1995) Montreal, Canada.
7]
R.C. GONZALEZ AND R.E. WOODS, Digital Image
Processing, Prentice Hall, 2002.
IBRAHIM S. I. ABUHAIBA is an associate professor at the Islamic University of Gaza, Department of Electrical and Computer Engineering,
Gaza, Palestine. He obtained his Master of Philosophy and Doctorate of Philosophy from Britain, in the field of document understanding
and pattern recognition. His research interests include computer vision, image processing, document analysis and understanding, pattern
recognition, artificial intelligence, and other fields. Dr. Abuhaiba presented important theorems and more than 30 algorithms in document
understanding. He published many original contributions in the field
of document understanding in well-reputed international journals and
conferences.